Frontier AI is fragmenting: Fable 5 is a cyber and systems specialist with glaring gaps, GPT‑5.6 looks like the new generalist, and Chinese MoE models are taking over the cost‑efficient coding tier just as geopolitics heats up. Generic enterprise copilots are largely failing in the wild while developers quietly shift to mix‑and‑match agent stacks, open‑weight coders, and aggressive systems hacks that slash $/token.
The real story is no longer who has the single 'smartest' model but who can juggle specialization, infra economics, and tightening governance without dropping anything.
Key Events
/Fable 5 was redeployed globally with a cybersecurity focus while routine coding tasks in its ecosystem default to Opus 4.8.
/The U.S. Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5, reopening international access.
/GPT‑5.6 Sol Ultra achieved 91.9% on TerminalBench 2.1, topping reported AI models.
/Open-source MoE GLM‑5.2 became the first open model to lead APEX‑SWE Integration with 55.3% Pass@1.
/DeepSeek V4 Flash can run locally with a 1M‑token context on an RTX 5090‑class GPU.
Report
Fable 5 now behaves less like a general upgrade and more like a cyber‑specialist: strong on labor automation and kernel benchmarks, weak on everyday debugging.
GPT‑5.6 Sol Ultra and open MoE models like GLM‑5.2 and DeepSeek V4 Flash are rising at the same time, so 'frontier' has quietly become a set of niches rather than a single winner.
frontier is splintering by capability
Fable 5 posts a 16.10% score on the Remote Labor Automation index, beating Opus on measured coverage of remote work tasks. It also wrote the fastest megakernel on KernelBench, with an 18.71× speedup over a PyTorch baseline after 2.5 hours of model time.
At the same time its debugging score fell from 86.2 to 25.9 on re‑evaluation, and routine coding in its own product stack is routed to older Opus 4.8.
By contrast GPT‑5.6 Sol Ultra is being positioned as a more classical 'do‑everything' upgrade, topping TerminalBench 2.1 at 91.9% and heading straight into Codex.
GLM‑5.2 runs as a ~750B‑parameter MoE with 40B active, became the first open model to lead APEX‑SWE Integration at 55.3% Pass@1, and is about 5× cheaper than Opus 4.8.
DeepSeek V4 Flash packs 284B parameters, offers 1M‑token context on an RTX 5090, and beats Sonnet and Opus on coding at lower cost.
Tencent’s Hy3 MoE is smaller in active parameters (21B) but roughly seven times cheaper than GLM‑5.2 while cutting hallucinations from 12.5% to 5.4%.
OpenRouter data shows U.S. companies already sending 46% of their tokens to Chinese models, even as Beijing considers restricting overseas access and the U.S. hires someone specifically to decide which models to ban.
On top of that, Anthropic accuses Alibaba’s Qwen team of a 28.8‑million‑interaction Claude distillation campaign, turning model choice into an IP and export‑control minefield as well as a price/perf decision.
copilots are flopping while dev agents fragment
Microsoft 365 Copilot has reached about 4.5% user engagement after three years, with only ~1% using it weekly, despite 6,000 engineers and price hikes justified by AI innovation.
Users routinely describe Copilot as ineffective even for simple tasks, and Microsoft is now replacing OpenAI and Anthropic models in Excel and Outlook primarily to cut Copilot costs.
Meanwhile GitHub Copilot quietly added Kimi K2.7 Code as its first open‑weight model option and re‑enabled Fable 5, signalling that even flagship tools want cheaper, swappable backends.
In the more technical crowd, Codex plus GPT‑5.6 is becoming the high‑end coding combo while GLM‑5.2‑powered ZCode, DeepSeek V4 Flash, and agents like Cline or Cursor compete on routing, context handling, and cost rather than a single monolithic IDE.
systems hacks are eating model economics
Anthropic cut Fable serving costs by about 60% not by changing the weights, but by rendering code as images and running it through OCR. The same stack had Fable 5 write a fused megakernel that ran 18.71× faster than a PyTorch baseline in just 2.5 hours of model time.
On the vision side, ComfyUI v0.27.0’s convrot INT8 models are more than 2× faster than fp16 on Nvidia 20–50 series GPUs, and Krea 2 Turbo pushes fully open 4K image generation, eroding the performance moat of proprietary media APIs.
For language, GLM‑5.2 in NVFP4 format hits around 24 tokens per second at 128K context on four DGX Sparks, while vLLM and DeepSeek’s DSpark scheduler push bulk captioning throughput toward 2000 transactions per second.
These are systems tricks—quantization formats, better kernels, smarter schedulers—that are moving $/token faster than raw scaling curves, quietly reshaping who can afford 'frontier' behavior.
agent stacks are powerful and brittle at the same time
In production, about 89% of agent teams have observability but only 52% actually evaluate their agents, which helps explain how a LangGraph system looped on a broken tool until it generated a massive API bill.
LiteLLM was picked for its simplicity, then saw costs unexpectedly triple over eight months as its community‑maintained price map drifted, forcing teams to add tagging and per‑key budgets after the fact.
OpenClaw, one of the more capable open agents, logged 495 CVEs in half a year, while many MCP servers ship without proper auth and are often blocked outright by enterprise security teams.
RAG pipelines keep gaining sophistication—with graph‑free multi‑hop setups and faithfulness regression gates—but still produce confident nonsense often enough that projects like Godot now ban AI‑generated code entirely and fanfic communities are building AI detectors to defend their norms.
What This Means
The center of gravity is drifting from 'pick the smartest model' to navigating a messy mix of specialized frontiers, geo‑risky MoEs, and fragile agent plumbing shaped by infra hacks and governance fights. The neat leaderboard story is giving way to a world where cost engineering, policy, and developer trust decide which capabilities actually reach users.
On Watch
/Whether U.S. and Chinese regulators actually follow through on model access controls, as Beijing weighs restricting overseas access to top models and the U.S. hires a dedicated official to decide which AIs to ban.
/How fast high‑end local inference catches up, with llama.cpp’s DFlash showing 4.44× speedups at 36K context, GLM‑5.2 reportedly running on a 25 GB‑RAM laptop, and Qualcomm’s GenieX driving 20 tok/s for Gemma 4 26B on Windows laptops.
/The hardening of norms around AI‑generated code, from Alibaba’s internal Claude Code ban to Godot rejecting AI patches and fanfic communities building AI detectors, which could snowball into de facto standards for provenance.
Interesting
/Sam Altman hinted at GPT-5.6's potential to discover new mathematical concepts, involving a Cambridge math student.
/Mistral's Leanstral-1.5-119B-A6B model excels in formal verification, achieving state-of-the-art results.
/Alibaba's DAMO Academy unveiled ElementsClaw, an AI agent that discovered four superconducting materials after screening 2.4 million crystal structures in 28 GPU hours.
/Speculative decoding on GPUs may lead to increased energy consumption per token, countering expectations of speed benefits.
/A new LLM engine has been created that can train on any GPU without relying on CUDA or PyTorch, available as a C++ binary.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/Fable 5 was redeployed globally with a cybersecurity focus while routine coding tasks in its ecosystem default to Opus 4.8.
/The U.S. Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5, reopening international access.
/GPT‑5.6 Sol Ultra achieved 91.9% on TerminalBench 2.1, topping reported AI models.
/Open-source MoE GLM‑5.2 became the first open model to lead APEX‑SWE Integration with 55.3% Pass@1.
/DeepSeek V4 Flash can run locally with a 1M‑token context on an RTX 5090‑class GPU.
On Watch
/Whether U.S. and Chinese regulators actually follow through on model access controls, as Beijing weighs restricting overseas access to top models and the U.S. hires a dedicated official to decide which AIs to ban.
/How fast high‑end local inference catches up, with llama.cpp’s DFlash showing 4.44× speedups at 36K context, GLM‑5.2 reportedly running on a 25 GB‑RAM laptop, and Qualcomm’s GenieX driving 20 tok/s for Gemma 4 26B on Windows laptops.
/The hardening of norms around AI‑generated code, from Alibaba’s internal Claude Code ban to Godot rejecting AI patches and fanfic communities building AI detectors, which could snowball into de facto standards for provenance.
Interesting
/Sam Altman hinted at GPT-5.6's potential to discover new mathematical concepts, involving a Cambridge math student.
/Mistral's Leanstral-1.5-119B-A6B model excels in formal verification, achieving state-of-the-art results.
/Alibaba's DAMO Academy unveiled ElementsClaw, an AI agent that discovered four superconducting materials after screening 2.4 million crystal structures in 28 GPU hours.
/Speculative decoding on GPUs may lead to increased energy consumption per token, countering expectations of speed benefits.
/A new LLM engine has been created that can train on any GPU without relying on CUDA or PyTorch, available as a C++ binary.